Representing $f$-divergences as solutions to optimisation problems
Avraham Ruderman (NICTA)
NICTA SML SEMINARDATE: 2011-08-25
TIME: 11:00:00 - 12:00:00
LOCATION: NICTA - 7 London Circuit
CONTACT: JavaScript must be enabled to display this email address.
ABSTRACT:
$f$-divergernces are a class of discrepancy measures between probability distributions. They generalise some well known discrepancy measures such as the KL-divergence, the variational divergence and the Hellinger distance. Recently there has been interest both in the theoretical as well as in the applied literature in representing $f$-divergences as solutions to optimisation problems. We show how the representations currently used in the literature can be tightened. We also show how one can simply derive an estimation procedure based on these representations. We also show a connection to MMD.
